How the study classifies a message as task crossover

To determine whether a message represents crossover work, two things must be established: what the user's role is, and what task the message describes. The report follows a four-step pipeline, illustrated by an example from Figure 1 of the report:

Click the four tabs above to step through the workflow. Re-created from Figure 1 in the report.

Two key concepts provide essential context. O*NET is the US Department of Labor's occupational database, breaking down professions into detailed work activities. The study uses it as the baseline standard for task ownership. A detailed work activity (DWA) is the smallest unit of measurement on this scale—such as "troubleshoot computer applications or systems" or "calculate financial data."

Role boundaries include a built-in buffer zone. The report does not enforce strict exact matches: if a task shares a semantic similarity score above 0.80 (out of 1.0) with the closest task in a worker's own occupational boundary, it is still classified as own-occupation work. As a result, the "own-occupation" threshold is generous, making the bar for task crossover higher than a literal reading suggests.

With the three categories defined, we reach the metric most distorted in media coverage.

The 43.5% figure comes from discarding 60% of messages

The 16.8% raw figure and the 43.5% headline figure are derived from the exact same pool of messages. The difference comes down to a single methodological step: dropping the 61.5% of generic work entirely from the denominator, leaving only the 38.5% of "specialized prompts," and re-basing that 38.5% to 100%. Calculating 16.8% over this reduced 38.5% denominator yields 43.5%.

Figure 2 from the original report. Top right specifies "denominator: all work-related messages"; middle notes "Retained 38.5%: non-generic"; bottom bar states "rebased to 100%." Source: OpenAI Work at the Frontier.

The report explicitly highlights this adjustment: the chart states "rebased to 100%," and the text specifies "excluding generic work." The nuance was lost primarily in secondary commentary.

The second statement sounds far more dramatic, yet the 60% it sets aside is precisely what most people use AI for every day: writing emails, editing copy, and scheduling meetings.

With the baseline defined, we can examine how the eight covered occupations compare under the report's specialized prompt framework. Unless noted otherwise, crossover ratios below refer to non-generic specialized prompts.

Customer experience and design cross boundaries most; engineering least

Breaking down the 43.5% figure across the eight functional roles reveals wide divergence: customer experience leads at 77%, while engineering sits at 28%—a nearly threefold difference.

Five of the eight roles exceed 50%. For workers in those top five dark-colored roles, when they ask ChatGPT a specialized question, more than half the time it falls outside their own domain.

Engineering's 28% figure needs context: it does not mean engineers avoid crossover work. In absolute volume, engineers generate vast amounts of own-occupation AI prompts—coding, debugging, code reviews—which fall squarely within engineering boundaries, expanding the denominator and diluting the crossover share. As shown later, crossover messages account for 18.5% of engineers' total work prompts (including generic work), also the lowest among the eight roles. That, however, is entirely separate from whether engineering tasks travel to other roles.

Marketing and engineering tasks travel to every other role

While the previous section shows which roles perform crossover work, Figure 4 from the report reveals whose tasks they are borrowing: each row represents a worker occupation, each column represents the traditional O*NET home of the task, and bubble size reflects percentage share.

Figure 4 from the report. Vertical axis shows worker occupation; horizontal axis shows traditional task source. Column headers: Customer experience, Design, Engineering, Finance, Human resources, Legal, Marketing, Sales. Dark diagonal cells indicate own-occupation tasks. Denominator is specialized non-generic prompts for each row; row totals may not sum to 100% due to rounding. Source: OpenAI Work at the Frontier.

The report notes that marketing and engineering tasks appear most frequently across other roles' AI usage. Specifically, marketing tasks account for 28%–29% of non-generic prompts from sales and design workers, and 26% from customer experience workers. Engineering tasks make up 28% of design workers' non-generic prompts, and 20%–22% for customer experience and finance workers.

Examining this matrix row by row reveals a key pattern not explicitly highlighted in the text:

Comparing rows reveals clear patterns summarized below:

Compiled from Figure 4 of the report. Labeled percentages per row sum to 82%–100% due to rounding and unlabeled minor bubbles. For finance, the 25% vs 23% gap is within margin of error.

Examining specific recurring tasks from Table 2 in the report highlights which activities travel most frequently across roles ("reach" indicates how many of the other seven occupations rank the task in their top three crossover items):

Financial data calculation and software troubleshooting show the broadest cross-role penetration, ranking in the top three across all seven other occupations. Among non-finance workers sending finance prompts, 9.8% focus on calculating financial data; among non-engineers sending engineering prompts, 6.1% involve software troubleshooting.

Among non-marketing workers generating marketing prompts, 25% focus on "developing promotional material"—ads, social posts, flyers, promo videos, decks, and product sheets—with 9.3% focused on "drafting marketing collateral" and 6.5% on "developing marketing strategies."

This raises a complementary question: which roles import tasks, and which roles export them?

Designers do everything, but nobody else does design

Figure 5 in the report separates task crossover into two directions: Tasks brought in measures how much external work a role incorporates into its AI prompts; Tasks that travel measures how often a role's traditional tasks appear in other workers' prompts.

Design occupies two opposite extremes: highest in tasks brought in (35.2%) and lowest in tasks that travel (1.7%)—a twentyfold imbalance. Designers leverage AI to absorb tasks from multiple functions, while non-designers rarely attempt design tasks via AI.

Engineering presents the reverse pattern: engineers import the fewest external tasks (18.5%), but engineering tasks travel second farthest (7.4%), making them among the most frequently borrowed by other functions.

Marketing leads on both fronts: 24.3% of marketing workers' prompts involve external tasks, while marketing tasks travel to other roles at a sample-wide high of 8.9%.

Figure 5 from the report. Left: Tasks brought in; Right: Tasks that travel. Both x-axes scale 0–40%. Source: OpenAI Work at the Frontier.

The report includes an explicit caveat: task crossover does not signal the disappearance of these roles. While AI lowers entry barriers for non-experts tackling basic tasks, domain experts remain essential for specialized judgment and oversight. The findings reflect evolving task allocation rather than direct forecasts of headcount changes.

Beyond individual roles, how does organization size affect task crossover?

Smaller workplaces show more crossover, but only among moderate users

Figure 6 groups workspaces by seat count: workspaces with 2–5 seats average 18.9% crossover prompts, compared to 16.3% for those with over 100 seats—appearing to show a neat downward trend.

However, three caveats apply:

First, the slope is modest. The vertical axis in Figure 6 spans only 16% to 19%—a 3 percentage point range. The 2.6 point shift represents a ~13% relative decline from small to large workspaces.

Second, the trend holds only for mid-tier users. When breaking down users by prompt volume (Appendix Figure A1), the pattern diverges across tiers.

Appendix Figure A1. Three panels show bottom 25%, middle 50%, and top 25% prompt volume tiers. Source: OpenAI Work at the Frontier.

Values in percentages. Data from Appendix Figure A1. The report notes that for the bottom 25% tier, the 101+ seat bin contains only 197 messages.

The main text cites the middle tier. For high-volume users (top 25%), values fluctuate without a clear monotonic trend—a detail noted in the report.

Third, workspace seat count differs from enterprise headcount. Organizations may purchase seats for only a subset of staff. Seat counts serve as proxy indicators for specialization, industry type, organizational maturity, or peer availability.

One hypothesis offered is that small firms lack dedicated specialists, forcing employees to handle marketing copy, troubleshooting, or contract reviews independently with AI. In larger enterprises, workers are more likely to delegate such tasks to specialized internal teams. This represents a plausible interpretation rather than a proven causal relationship.

Crucially, all these proportions depend on excluding generic work. Restoring generic tasks alters the picture significantly.

Including generic work shows core domain retention across all roles

On page 15, the report re-calculates the matrix after assigning generic work back to each occupation's own category:

Appendix Figure A2. Subtitle notes "Dark diagonal = own-occupation + shared work." Diagonal values top-to-bottom: Customer experience 70%, Design 65%, Engineering 82%, Finance 70%, Human resources 70%, Legal 75%, Marketing 76%, Sales 68%. Source: OpenAI Work at the Frontier.

All eight occupations fall between 65% and 82%. Customer experience (77% crossover in specialized prompts) sits at 70% core retention here, while design (75% crossover) sits at 65%.

Task crossover does not replace role-specific work: every occupation still retains a meaningful diagonal of same-occupation or shared activity.

Simultaneously, non-diagonal matrix entries show workers regularly combining own-role duties with adjacent tasks—particularly in marketing and engineering. The report characterizes this as expanding job scope and blending task mixes, rather than wholesale role replacement.

This contextual adjustment appears on page 15, twelve pages after the headline figures on page 3.

Two core questions raised by the report

If core responsibilities remain intact, where do the practical impacts lie? The conclusion highlights two structural questions without offering definitive answers.

First: If workers routinely use AI to perform tasks outside their domain, will this evolve into permanent job role expansion? If so, organizations must train staff to evaluate AI outputs outside their expertise and establish clear verification and governance workflows.

Consider practical scenarios: When a marketing specialist uses AI to patch website code, who verifies the fix? When a customer service rep drafts a regulatory response using AI, who signs off? In Table 2, 21.4% of regulatory communication tasks originated from CX prompts. The report highlights this operational challenge without prescribing solutions.

Second: Official labor databases log job duties based on historical standards. As AI alters task distribution, classification frameworks like O*NET risk falling out of step with actual workplace practices.

Historically, task blurring precedes new job titles

A key observation in the background section highlights why this dataset is significant:

Analysis of Generative AI usage offers a window into an earlier stage of the job evolution process, as workers can begin testing and recombining tasks before firms revise job descriptions or invent new job titles altogether.

Work at the Frontier, p. 4

Traditional economic research tracks job evolution long after changes surface in hiring ads, job titles, and labor statistics. This prompt data captures an earlier phase: a designer prompting AI late at night for an engineering solution—an interaction invisible to HR records.

The three papers cited in the report connect this timeline:

All three studies are cited in the report's background section. Co-author Caroline Chin also co-authored New Frontiers: The Origins and Content of New Work, 1940–2018.

Together, these studies contextualize the report: occupational boundaries have been evolving for decades, and prompt data captures a real-time snapshot of that ongoing shift.

How changes may surface first

The following scenario represents an analytical synthesis based on report data rather than an explicit report forecast.

As crossover persists, initial shifts will likely appear not in formal job titles, but in day-to-day requirements. Preferred qualifications in job postings will expand—a customer support role seeking basic SQL literacy, or a design role seeking analytics comprehension. Performance evaluations will encounter blurred attribution for shared deliverables. Training budgets will partially pivot from deep domain specialization toward verifying cross-functional AI outputs.

Accountability will present the primary friction. When a customer service employee uses AI to draft a regulatory response, does legal review every submission? Mandating review recreates bottlenecks; skipping review creates risk. The report identifies this operational dilemma, which organizations must navigate through internal process design.

While headlines highlight 43.5%, the report also details a 16.8% raw figure. The difference stems entirely from denominator selection: filtering out the 61.5% share of generic work (emails, meetings, editing) and re-basing the remaining 38.5% specialized prompts to 100% transforms 16.8% into 43.5%.

The exact same crossover data point (animated dot above) shifts from 16.8% to 43.5% based on denominator selection. The report clearly indicates "rebased to 100%" and "excluding generic work"; secondary coverage omitted this context.

When evaluated strictly on specialized prompts, task crossover exceeds 50% in five of eight occupations, with nearly a threefold gap between the highest and lowest roles.

For the top five dark-colored roles, specialized AI prompts focus more often on external tasks than own-domain work. Engineering's 28% figure reflects high internal coding prompt volume expanding its denominator rather than a lack of task crossover.

Task crossover operates in two directions: Tasks brought in (external tasks in role prompts) and Tasks that travel (role tasks in external prompts).

Both sides drawn to a common scale (max 40%). Three patterns emerge: Design integrates work from multiple domains while exporting very little; Engineering retains strong internal focus while exporting tasks outward; Marketing sees high activity in both directions.

When generic work is reassigned to each role's own baseline, 65%–82% of all activity across all eight occupations remains centered on own-domain and shared work (Engineering 82% max, Design 65% min). Task crossover adds to existing roles rather than replacing them.